Combining Resnet 34 with U-net for the Segmentation of retinal blood vessels

Abstract

Blood vessels eye in the human body give us important information about a person's health, it allows us to identify diseases. Blood vessel segmenta-tion facilitates these operations. In deep learning, especially when convolutional layers are used more frequently within the training model, there is deterioration and divergence from the best results. In this paper, we have treated and solved the gradient problem by proposing a deep learning architecture for the segmen-tation of the vascular networks of blood vessels in fundus images. This architec-ture combines residual learning and U-Net. As we know U-net, it consists of two parts, coder and decoder paths, we combined Resnet 34 with U-net as follows. In downsampling where we are extracted features of the image by convolution with a filter in every layer that we define. Here and at this stage, we used Resnet 34 for extracted our features. After extraction of features in the step of downsampling, we go to a stage upsampling with the same number of layers that we used in downsampling, and of course, we apply a concatenation between the patches. Finally٫ we got the value of recall equal to 0.9794, Accuracy 0.9692, sensitivity equal to 0.7859, specificity equal to 0.9870 and 0.9832 F1-Score for DRIVE (Digital Retinal Images for Vessels Extraction) database, and with STARE (Structuring Analysis of the Retina) database, we got recall equal 0.9961, Accuracy 0.9363, sensitivity equal 0.9335, plus specificity equal 0.9246 and 0.9649 F1-Score. We can apply our model for the segmentation of vessels similar to different elements in the medical field. This work outperforms many previous contributions.